doctr.transforms

Data transformations are part of both training and inference procedure. Drawing inspiration from the design of torchvision, we express transformations as composable modules.

Supported transformations

Here are all transformations that are available through docTR:

class doctr.transforms.modules.Resize(output_size: int | Tuple[int, int], method: str = 'bilinear', preserve_aspect_ratio: bool = False, symmetric_pad: bool = False)

Resizes a tensor to a target size

>>> import tensorflow as tf
>>> from doctr.transforms import Resize
>>> transfo = Resize((32, 32))
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Parameters:

• output_size – expected output size

• method – interpolation method

• preserve_aspect_ratio – if True, preserve aspect ratio and pad the rest with zeros

• symmetric_pad – if True while preserving aspect ratio, the padding will be done symmetrically

class doctr.transforms.modules.GaussianNoise(mean: float = 0.0, std: float = 1.0)

Adds Gaussian Noise to the input tensor

>>> import tensorflow as tf
>>> from doctr.transforms import GaussianNoise
>>> transfo = GaussianNoise(0., 1.)
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Parameters:

• mean – mean of the gaussian distribution

• std – std of the gaussian distribution

class doctr.transforms.modules.ChannelShuffle

Randomly shuffle channel order of a given image

class doctr.transforms.modules.RandomHorizontalFlip(p: float)

Adds random horizontal flip to the input tensor/np.ndarray

>>> import tensorflow as tf
>>> from doctr.transforms import RandomHorizontalFlip
>>> transfo = RandomHorizontalFlip(p=0.5)
>>> image = tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1)
>>> target = np.array([[0.1, 0.1, 0.4, 0.5] ], dtype= np.float32)
>>> out = transfo(image, target)

Parameters:

p – probability of Horizontal Flip

class doctr.transforms.modules.RandomShadow(opacity_range: Tuple[float, float] | None = None)

Adds random shade to the input image

>>> import tensorflow as tf
>>> from doctr.transforms import RandomShadow
>>> transfo = RandomShadow(0., 1.)
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Parameters:

opacity_range – minimum and maximum opacity of the shade

class doctr.transforms.modules.RandomResize(scale_range: Tuple[float, float] = (0.3, 0.9), preserve_aspect_ratio: bool | float = False, symmetric_pad: bool | float = False, p: float = 0.5)

Randomly resize the input image and align corresponding targets

>>> import tensorflow as tf
>>> from doctr.transforms import RandomResize
>>> transfo = RandomResize((0.3, 0.9), preserve_aspect_ratio=True, symmetric_pad=True, p=0.5)
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Parameters:

• scale_range – range of the resizing factor for width and height (independently)

• preserve_aspect_ratio – whether to preserve the aspect ratio of the image,

• value (given a float)

• probability (the symmetric padding will be applied with this)

• symmetric_pad – whether to symmetrically pad the image,

• value

• probability

• p – probability to apply the transformation

Composing transformations

It is common to require several transformations to be performed consecutively.

class doctr.transforms.modules.SampleCompose(transforms: List[Callable[[Any, Any], Tuple[Any, Any]]])

Implements a wrapper that will apply transformations sequentially on both image and target

PyTorchTensorFlow

>>> import numpy as np
>>> import torch
>>> from doctr.transforms import SampleCompose, ImageTransform, ColorInversion, RandomRotate
>>> transfos = SampleCompose([ImageTransform(ColorInversion((32, 32))), RandomRotate(30)])
>>> out, out_boxes = transfos(torch.rand(8, 64, 64, 3), np.zeros((2, 4)))

Parameters:

transforms – list of transformation modules

class doctr.transforms.modules.ImageTransform(transform: Callable[[Any], Any])

Implements a transform wrapper to turn an image-only transformation into an image+target transform

PyTorchTensorFlow

>>> import torch
>>> from doctr.transforms import ImageTransform, ColorInversion
>>> transfo = ImageTransform(ColorInversion((32, 32)))
>>> out, _ = transfo(torch.rand(8, 64, 64, 3), None)

Parameters:

transform – the image transformation module to wrap

class doctr.transforms.modules.ColorInversion(min_val: float = 0.5)

Applies the following tranformation to a tensor (image or batch of images): convert to grayscale, colorize (shift 0-values randomly), and then invert colors

PyTorchTensorFlow

>>> import torch
>>> from doctr.transforms import ColorInversion
>>> transfo = ColorInversion(min_val=0.6)
>>> out = transfo(torch.rand(8, 64, 64, 3))

Parameters:

min_val – range [min_val, 1] to colorize RGB pixels

class doctr.transforms.modules.OneOf(transforms: List[Callable[[Any], Any]])

Randomly apply one of the input transformations

PyTorchTensorFlow

>>> import torch
>>> from doctr.transforms import OneOf
>>> transfo = OneOf([JpegQuality(), Gamma()])
>>> out = transfo(torch.rand(1, 64, 64, 3))

Parameters:

transforms – list of transformations, one only will be picked

class doctr.transforms.modules.RandomApply(transform: Callable[[Any], Any], p: float = 0.5)

Apply with a probability p the input transformation

PyTorchTensorFlow

>>> import torch
>>> from doctr.transforms import RandomApply
>>> transfo = RandomApply(Gamma(), p=.5)
>>> out = transfo(torch.rand(1, 64, 64, 3))

Parameters:

• transform – transformation to apply

• p – probability to apply

class doctr.transforms.modules.RandomRotate(max_angle: float = 5.0, expand: bool = False)

Randomly rotate a tensor image and its boxes

Parameters:

• max_angle – maximum angle for rotation, in degrees. Angles will be uniformly picked in [-max_angle, max_angle]

• expand – whether the image should be padded before the rotation

class doctr.transforms.modules.RandomCrop(scale: Tuple[float, float] = (0.08, 1.0), ratio: Tuple[float, float] = (0.75, 1.33))

Randomly crop a tensor image and its boxes

Parameters:

• scale – tuple of floats, relative (min_area, max_area) of the crop

• ratio – tuple of float, relative (min_ratio, max_ratio) where ratio = h/w